Multi-agent budgets must apply to the shared task, including retries and delegation. Per-agent limits can still permit expensive loops when each handoff starts a fresh allowance.
Budget control belongs to the orchestrator
A multi-agent budget limits included resource use and effects across a logical task. It can cover model calls, tool operations, external messages, elapsed time, and delegated depth. Define the units explicitly.
For a proposed controller, reserve or account for work under a trusted task identity. Child agents inherit the relevant allowance and cannot reset it through a new message. Track progress in durable state rather than accepting generated claims that more work is necessary.
Keep spending estimates separate from settled usage where provider accounting is delayed. A budget based only on late invoices may not contain an active loop in time.
Reproduce a fictional delegation cycle
Agent A asks B to validate a service rule. B sends the unresolved task back to A with a new phrasing. Each sees a fresh request and continues. The orchestrator detects repeated ownership transitions without new relevant evidence and stops the cycle under its policy.
A second fixture makes a tool fail repeatedly. Retries consume the shared allowance and obey a time budget. The task should stop with a known unresolved state before uncontrolled resource use or repeated external effects occur.
| Budget fixture | Required control |
|---|---|
| Circular handoff | Bounded delegation and no-progress detection |
| New child agent | Inherits parent task scope |
| Repeated tool failure | Shared attempt and time limits |
| Concurrent work claims | Atomic allowance accounting |
| Budget reached during uncertainty | Preserve reconciliation and next owner |
Inspect actual usage and destination effects in a fake harness. A controller reporting “within budget” is insufficient if workers can bypass its accounting path.
Stop usefully rather than silently
The terminal record should show the completed work, unresolved question, known side effects, reason for stopping, and approved continuation path. A budget stop should not be misreported as successful task completion.
Tight budgets can interrupt legitimate hard tasks. Use observed requirements to set limits and permit explicit authorised expansion where appropriate. Keep cleanup and reconciliation governed so containment does not abandon consequential effects halfway through.
This controller is proposed. The voice integration case study provides related context. Pair it with runaway automation containment.
Multi-agent budget FAQ
Can every agent have an independent budget?
Local limits can help, but enforce the aggregate task boundary too.
Should handoff count as progress?
Only when it advances a defined task state or obtains relevant new evidence.
What should the caller or staff hear at the cap?
An accurate unresolved status and available next step, not a fabricated completion or unexplained disappearance.